Troubleshooting

General Usage

  • Problem: Issues with the +cpu PyTorch package.

    • Cause: Certain Python packages may have PyTorch as a hard dependency. If you installed the +cpu version of PyTorch, installation of these packages might replace the +cpu version with the default version released on Pypi.org.

    • Solution: Reinstall the +cpu version back.

  • Problem: The workload running with Intel® Extension for PyTorch* occupies a remarkably large amount of memory.

    • Solution: Try to reduce the occupied memory size by setting the --weights_prepack parameter of the ipex.optimize() function to False.

  • Problem: The conv+bn folding feature of the ipex.optimize() function does not work if inference is done with a custom function:

    import torch
    import intel_pytorch_extension as ipex
    
    class Module(torch.nn.Module):
        def __init__(self):
            super(Module, self).__init__()
            self.conv = torch.nn.Conv2d(1, 10, 5, 1)
            self.bn = torch.nn.BatchNorm2d(10)
            self.relu = torch.nn.ReLU()
    
        def forward(self, x):
            x = self.conv(x)
            x = self.bn(x)
            x = self.relu(x)
            return x
    
        def inference(self, x):
            return self.forward(x)
    
    if __name__ == '__main__':
        m = Module()
        m.eval()
        m = ipex.optimize(m, dtype=torch.float32, level="O0")
        d = torch.rand(1, 1, 112, 112)
        with torch.no_grad():
          m.inference(d)
    
    • Cause: PyTorch FX limitation.

    • Solution: You can avoid this error by calling m = ipex.optimize(m, level="O0"), which doesn’t apply ipex optimization, or disable conv+bn folding by calling m = ipex.optimize(m, level="O1", conv_bn_folding=False).

Performance Regression

  • Some models may experience performance regression comparing to 2.0.x due to deprecation of the NNC feature in PyTorch*.

TorchDynamo

  • Problem: A workload that uses torch.compile() fails to run or demonstrates poor performance.

    • Cause: The support of torch.compile() with ipex as the backend is still an experimental feature. Currently, the following HuggingFace models fail to run using torch.compile() with ipex backend due to memory issues:

      • masked-language-modeling+xlm-roberta-base

      • casual-language-modeling+gpt2

      • casual-language-modeling+xlm-roberta-base

      • summarization+t5-base

      • text-classification+allenai-longformer-base-409

    • Solution: Use the torch.jit APIs and graph optimization APIs of the Intel® Extension for PyTorch*.

Dynamic Shape

  • Problem: When working with an NLP model inference with dynamic input data length using TorchScript (either torch.jit.trace or torch.jit.script), performance with Intel® Extension for PyTorch* may be less than that without Intel® Extension for PyTorch*.

    • Solution: Use the workaround below:

      • Python interface

        torch._C._jit_set_texpr_fuser_enabled(False)
        
      • C++ interface

        #include <torch/csrc/jit/passes/tensorexpr_fuser.h>
        torch::jit::setTensorExprFuserEnabled(false);
        

INT8

  • Problem: Limitations of dynamic shapes support of static quantization:

    • When an input shape is provided in runtime for the first time, execution could take longer time to compile a new kernel for this shape. Specifically, the new kernel compilation time could be long for complicated kernels.

    • Channels Last format won’t take effect with dynamic input shapes for CNN models at this time. Optimizations are undergoing.

  • Problem: RuntimeError: Overflow when unpacking long when a tensor’s min max value exceeds int range while performing int8 calibration.

    • Solution: Customize QConfig to use min-max calibration method.

  • Problem: Models get large accuracy loss with the default quantization recipe.

  • Problem: Incorrect results with large tensors when calibrating with quantize_per_tensor, when benchmarking with 1 OpenMP* thread (find more detailed info here.

    • Solution: Editing your code following the pseudocode below can workaround this issue, if you do need to explicitly set OMP_NUM_THREAEDS=1 for benchmarking. However, there could be a performance regression if oneDNN graph compiler prototype feature is used.

      Workaround pseudocode:

      # perform convert/trace/freeze with omp_num_threads > 1(N)
      torch.set_num_threads(N)
      prepared_model = prepare(model, input)
      converted_model = convert(prepared_model)
      traced_model = torch.jit.trace(converted_model, input)
      freezed_model = torch.jit.freeze(traced_model)
      # run freezed model to apply optimization pass
      freezed_model(input)
      
      # benchmarking with omp_num_threads = 1
      torch.set_num_threads(1)
      run_benchmark(freezed_model, input)
      
  • For models with dynamic control flow, please try dynamic quantization. Users are likely to get performance gain for GEMM models.

  • Support for EmbeddingBag with INT8 when bag size > 1 is work in progress.

BFloat16

  • Problem: BF16 AMP(auto-mixed-precision) runs abnormally with the extension on the AVX2-only machine if the topology contains Conv, Matmul, Linear, and BatchNormalization.

    • Solution: TBD

Runtime Extension

The following limitations currently exist:

  • Runtime extension of MultiStreamModule does not support DLRM inference, since the input of DLRM (EmbeddingBag specifically) cannot be simply batch split.

  • Runtime extension of MultiStreamModule has poor performance of RNNT Inference comparing with native throughput mode. Only part of the RNNT models (joint_net specifically) can be jit traced into graph. However, in one batch inference, joint_net is invoked multiple times. It increases the overhead of MultiStreamModule as input batch split, thread synchronization and output concat.

Result Correctness

  • Problem: Incorrect Conv and Linear result if the number of OMP threads is changed at runtime.

    • Cause: The oneDNN memory layout depends on the number of OMP threads, which requires the caller to detect the changes for the # of OMP threads while this release has not implemented it yet.

Float32 Training

  • Problem: Low throughput with DLRM FP32 Train.

    • Solution: A ‘Sparse Add’ PR is pending on review. The issue will be fixed when the PR is merged.